v2.5.1: Slash Commands Work Again on Custom LLMs, and Kanban Badges Jump Straight to the Card

If you run one of your AI employees on a custom LLM, you may have noticed something annoying lately: typing /compact did nothing, and switching models only changed a badge without changing the model. This release is mostly about that, and the fix turned out to be a single, satisfying root cause.

Slash commands reach the CLI again

Every message your employee receives gets a small speaker prefix stapled to the front, the [Message from ... ] marker that tells the model who is talking. Useful for chat. A problem for commands.

Because that prefix was being glued onto every input, your slash commands were arriving mangled. The CLI checks for an exact match, it looks for text === "/compact", and [Message from you] /compact is not that. So /compact, /model, /models, and /effort all sailed past the command handler and got treated as ordinary chat messages. The model would sometimes even reply to /compact as if you had asked it a question.

The fix is one guard in the right place. The admin proxy now never prefixes an input that starts with a / (a slash command) or a { (a JSON control frame). Everything else still gets its speaker prefix. That one change restores slash commands across every custom-LLM, Pi, and Claude session at once.

Compaction actually compacts now

With commands flowing through again, /compact does what it should on both engines.

On Pi, /compact now runs Pi's native compaction instead of being ignored or echoed back at you as text. On Claude, it is handed through as a real headless slash command, on top of Claude's own automatic compaction. Either way, when your conversation is getting heavy, you can trim it on demand.

We also killed a hang that made compaction feel broken even when it worked. A Pi compaction is not a normal agent turn, so nothing was telling the interface it had finished. The "Compacting..." banner could sit there forever, Stop button stuck, typing dots spinning. Now a successful Pi compaction posts its result as a collapsible "Conversation Summary" card and emits a proper stopped signal, so the banner, the Stop button, and the typing dots all clear exactly like the end of a normal turn.

One honest note on scope: if your conversation is genuinely small, under Pi's keep-recent window, there is nothing to compact and no summary appears. That is expected. The Compact button only shows up once you are at fifty percent context or more, where there is always enough to work with.

Switching a Claude model, live

Claude runs headless through a bridge, and a typed /model X on that bridge was being quietly dropped. The bridge only listens for a couple of specific message types, so your model switch changed the badge in the interface and nothing underneath. You thought you were on the new model. You were not.

The switch now goes through Claude's real set_model control request, the same channel the official SDK uses. It happens live, with no session respawn, so you can move between models mid-conversation and the change actually takes.

Kanban badges jump straight to the card

Small one, but it removes a daily papercut. When a Kanban notification shows up in your chat ("Kanban: card title created by you"), clicking it used to open the board and leave you to hunt for the card. Now it opens the board, switches to the right column, and pops the card's edit dialog open, exactly as if you had clicked the card yourself.

A quieter fix that saves a broken website

We also tightened the guidance baked into the web-publish skill. When your employee ships a live site, it should publish from a dedicated folder at the root of your data directory, never from the temp folder. Temp is for throwaway files and gets cleaned up on a schedule, which means a site published from there would quietly 404 later. New employees now inherit this best practice from the template, so a site that works today keeps working next week.

That is v2.5.1. No new headline feature, just your commands landing where they are supposed to, your compaction finishing cleanly, your model switches taking hold, and one fewer click on the board.

Want to test the most advanced AI employees? Try it here: https://Geta.Team